# Reasoning Model

AI Glossary · LLMs & Generative AI
Canonical URL: https://aisat.uk/glossary/reasoning-model
Author: Lucy Gilmore, Assistant General Counsel, Product, Privacy & AI at Sprinklr | Practical Privacy, Product & AI Governance & Defensible Risk Decisions (https://www.linkedin.com/in/lucy-gilmore/)
Breadcrumb: [Glossary](https://aisat.uk/glossary) > [LLMs & GenAI](https://aisat.uk/glossary?category=llms) > Reasoning Model

## TL;DR

> **A reasoning model** is a language model trained with reinforcement learning to generate and verify an internal chain of thought before producing a final answer. By spending extra compute at inference time, it excels at complex mathematics, multi-step logic, architecture planning, and code debugging.

## How Reasoning Model works in practice

Standard language models emit their answer token by token in a single forward pass, spending roughly the same amount of compute whether asked for the capital of France or asked to debug a race condition across four microservices. Reasoning models, such as OpenAI's o-series and GPT-5 reasoning modes, Anthropic's extended thinking models, and DeepSeek-R1, introduce a second scaling axis: test-time compute.

These models are post-trained using Reinforcement Learning with Verifiable Rewards (RLVR) against millions of problems with objective ground truth, such as unit-tested software patches, formal logic puzzles, and mathematical proofs. During training, the model learns to decompose problems into sub-goals, test hypotheses, spot dead ends, backtrack, and verify constraints inside hidden or inspectable 'thinking' tokens before emitting the visible response. Developers can typically configure a reasoning effort parameter (`low`, `medium`, `high`) or a thinking-token budget per API call.

For example, a Cambridge biotech firm uses a reasoning model to audit Python bioinformatics pipelines and design multi-step SQL cohort queries across clinical trial tables, where a single logical slip in an outer join would invalidate a study cohort.

## Why it matters and trade-offs

Reasoning models solve tasks where first-instinct pattern matching fails: complex tax or regulatory reconciliation, multi-file refactoring, security vulnerability analysis, and orchestrator planning in agentic workflows. On hard coding and scientific benchmarks, test-time thinking regularly beats a tenfold larger standard model.

The trade-offs are latency, token cost, and overthinking on simple tasks. Because reasoning tokens are billed as output tokens and generated sequentially, a single query can take ten to forty seconds and cost five to ten times more than a non-reasoning completion. Production systems should use model routing: send fast chat, classification, and simple RAG synthesis to low-latency models, and escalate only ambiguous planning or verification steps to a reasoning model.

## Related AI concepts

- [Large Language Model (LLM)](https://aisat.uk/glossary/large-language-model): A Large Language Model (LLM) is a deep transformer neural network containing billions of parameters trained on massive text and code corpora to comprehend, summarise, translate, reason over and generate human and programming languages.
- [Agentic AI](https://aisat.uk/glossary/agentic-ai): Agentic AI refers to artificial intelligence systems engineered to pursue multi-step objectives with bounded autonomy by decomposing goals, invoking external tools, inspecting intermediate results and adapting plans under defined governance controls.
- [Evals (Evaluations)](https://aisat.uk/glossary/evals): Evals (Evaluations) are automated, repeatable testing pipelines that score an AI application's outputs against a golden dataset using deterministic assertions, reference metrics and LLM-as-a-judge rubrics before any prompt or model change ships.
- [Small Language Model (SLM)](https://aisat.uk/glossary/small-language-model): A Small Language Model (SLM) is a compact neural language model, typically between one billion and ten billion parameters, optimised for low-latency, low-cost inference on a single GPU or on-device NPU for domain-specific workflows.